Library — v1.0

AI UX Design Patterns

A working library of AI design patterns for teams shipping AI products. Each pattern bridges one of the six HAX principles to a concrete piece of UI you can steal on Monday. Examples cite real products; anti-patterns are the mistakes we keep seeing in audits.

Missing a pattern? Send it to info@hax.academy — we update this page monthly.

Production examples are sourced from the aiverse.design pattern library.

Empathy First

Design for the human situation before the model capability.

Progressive Intent Capture

Problem: A blank prompt box forces users to invent structure they don't have.
Solution: Ask 2–3 short, sequential questions that shape the prompt behind the scenes.
Example: Notion AI's 'Continue writing / Summarize / Brainstorm' pre-intents before the free-text field.
Anti-pattern: One giant textarea labeled 'Ask anything'.

In production — via aiverse.design

Claude Code — Claude Code's autosuggesting promptsWatch on aiverse.design ↗

Grounded Empty States

Problem: First-run screens with a blinking cursor and no worked example.
Solution: Ship 3–5 tappable example prompts drawn from real user tasks, not marketing copy.
Example: Perplexity's category chips on the home canvas.
Anti-pattern: A hero image and a subtitle like 'The future of X'.

In production — via aiverse.design

Google Chrome — Google Chrome's personalized suggestionsWatch on aiverse.design ↗

Right-Sized Autonomy

Automation vs. augmentation is a per-step decision, not a product-wide toggle.

Human-in-the-Loop Gate

Problem: Agents that act without a checkpoint burn user trust on the first mistake.
Solution: Insert an explicit approve / edit / reject step before any irreversible action.
Example: Cursor's 'Apply' diff before code is written to disk.
Anti-pattern: 'Auto-run' as the only mode.

In production — via aiverse.design

Notion — Notion's verifying AI updates (diff)Watch on aiverse.design ↗

Autonomy Slider

Problem: One-size 'AI mode' ignores that users want more control on some tasks.
Solution: Expose Manual → Suggest → Draft → Auto as a first-class control per task.
Example: Linear's 'Draft with AI' vs. Gmail's Smart Compose inline ghost text.
Anti-pattern: Global on/off in settings, three menus deep.

In production — via aiverse.design

Gmail — Gmail's email predict textWatch on aiverse.design ↗

Transparent Capability

Say what the system can and cannot do, at the moment of use.

Capability Card

Problem: Users over-trust or under-trust because scope is invisible.
Solution: Inline card near input: what the model reads, what it doesn't, freshness of data, known limits.
Example: GitHub Copilot Chat's 'References' panel showing which files it saw.
Anti-pattern: A single 'AI can make mistakes' disclaimer in 10px grey.

In production — via aiverse.design

Cursor — Cursor's inline file referencingWatch on aiverse.design ↗

Calibrated Confidence Signal

Problem: Every answer looks equally certain, even the hallucinated ones.
Solution: Differentiate outputs the model is grounded on vs. inferring — badges, sourcing, or a low-confidence banner.
Example: Perplexity's per-sentence citations and 'Answer' vs. 'Steps' modes.
Anti-pattern: Bolded prose with no source anywhere.

In production — via aiverse.design

Gemini — Gemini's double check responseWatch on aiverse.design ↗

Real Control & Editability

AI output should never feel final.

Inline Refinement

Problem: Users have to rewrite the whole prompt to change one word.
Solution: Select-to-edit on any span: 'shorter', 'more formal', 'cite source'.
Example: Canva Magic Edit — apply, revert, or refine a single object.
Anti-pattern: Regenerate button that rerolls the entire output.

In production — via aiverse.design

Canva — Canva's ImagesWatch on aiverse.design ↗

Reversible Versions

Problem: AI overwrites the user's work with no undo across sessions.
Solution: Persistent version history with side-by-side diff and one-click revert.
Example: Google Docs' version history extended to AI edits.
Anti-pattern: Ctrl+Z as the only safety net.

In production — via aiverse.design

Zed.Dev — Zed.dev's granular code editingWatch on aiverse.design ↗

Graceful Failure

Failure is a design surface, not an exception.

Explain, Then Recover

Problem: Errors like 'Something went wrong' leave users guessing.
Solution: Name what failed, why the model thinks it failed, and offer 2 concrete recoveries.
Example: Claude's 'I can't do X because Y — try Z' redirects.
Anti-pattern: A red toast and a spinning retry.

In production — via aiverse.design

Claude — Claude's 404Watch on aiverse.design ↗

Graceful Refusal

Problem: Blanket refusals feel arbitrary and hostile.
Solution: State the policy, the closest thing the model *can* do, and a link to escalate.
Example: ChatGPT's 'I can't help with X, but I can help you draft a policy about it'.
Anti-pattern: 'As an AI language model, I cannot…'.

In production — via aiverse.design

Claude — Claude's spam disclaimerWatch on aiverse.design ↗

Mental Models

People don't have a mental model for AI yet. Your product gives them one.

Cheap Feedback

Problem: Thumbs-up/down with no follow-up teaches the model nothing useful.
Solution: One-tap signal + optional 'what was off?' chips (too long, wrong tone, hallucinated).
Example: Midjourney's variation/upscale buttons as implicit preference signals.
Anti-pattern: A modal survey after every response.

In production — via aiverse.design

Claude Code — Claude code's feedback collectionWatch on aiverse.design ↗

Visible Memory

Problem: Users don't know what the system remembers, so they can't trust it.
Solution: A dedicated 'Memory' surface: view, edit, delete individual facts the model retains.
Example: ChatGPT's Memory pane; Notion's per-workspace AI context.
Anti-pattern: Silent personalization with no export or delete.

In production — via aiverse.design

ChatGPT — ChatGPT's memory managementWatch on aiverse.design ↗

Want these patterns applied to your product?

The full-day HAX workshop audits a real AI product against the 7 Sins, then redesigns it using this pattern library.

See the workshop →